ARTICLE

DOI: 10.1038/s41467-018-06788-9 OPEN Patterns and drivers of recent disturbances across the temperate biome

Andreas Sommerfeld 1, Cornelius Senf 1,2, Brian Buma 3, Anthony W. D’Amato 4, Tiphaine Després 5,6, Ignacio Díaz-Hormazábal7, Shawn Fraver8, Lee E. Frelich 9, Álvaro G. Gutiérrez 7, Sarah J. Hart10, Brian J. Harvey11, Hong S. He12, Tomáš Hlásny5, Andrés Holz 13, Thomas Kitzberger14, Dominik Kulakowski 15, David Lindenmayer 16, Akira S. Mori17, Jörg Müller18,19, Juan Paritsis 14, George L. W. Perry 20, Scott L. Stephens21, Miroslav Svoboda5, Monica G. Turner 22, Thomas T. Veblen23 & Rupert Seidl 1 1234567890():,;

Increasing evidence indicates that forest disturbances are changing in response to global change, yet local variability in disturbance remains high. We quantified this considerable variability and analyzed whether recent disturbance episodes around the globe were con- sistently driven by , and if human influence modulates patterns of forest disturbance. We combined remote sensing data on recent (2001–2014) disturbances with in-depth local information for 50 protected landscapes and their surroundings across the temperate biome. Disturbance patterns are highly variable, and shaped by variation in disturbance agents and traits of prevailing species. However, high disturbance activity is consistently linked to warmer and drier than average conditions across the globe. Disturbances in protected areas are smaller and more complex in shape compared to their surroundings affected by human . This signal disappears in areas with high recent natural disturbance activity, underlining the potential of climate-mediated disturbance to transform forest landscapes.

1 University of Natural Resources and Sciences (BOKU) Vienna, Institute of Silviculture, Peter Jordan Straße 82, 1190 Wien, Austria. 2 Geography Department, Humboldt-Universität zu Berlin, Unter den Linden 6, 10099 Berlin, Germany. 3 Dept. of Integrative , University of Colorado, 1151 Arapahoe, Denver, CO 80204, USA. 4 University of , Rubenstein School of Environment and Natural Resources, Aiken Center Room 204E, Burlington, VT 05495, USA. 5 Faculty of and Wood Sciences, Czech University of Life Sciences in Prague, Kamýcká 129, 165 21 Prague 6, Czech Republic. 6 Institut de Recherche sur les Forêts, Université du Québec en Abitibi-Témiscamingue, 445 boulevard de l’Université, Rouyn-Noranda, QC J9X 5E4, Canada. 7 Facultad de Ciencias Agronómicas, Departamento de Ciencias Ambientales y Recursos Naturales Renovables, Universidad de Chile, Av. Santa Rosa 11315, La Pintana, 8820808 Santiago, Chile. 8 University of , School of Forest Resources, 5755 Nutting Hall, Orono, Maine 04469, USA. 9 Department of Forest Resources, University of Minnesota, 1530 Cleveland Ave. N., St.Paul, MN 55108, USA. 10 Department of Forest and Wildlife , University of Wisconsin–Madison, Madison, WI 53706, USA. 11 School of Environmental and Forest Sciences, University of Washington, Seattle, WA 98195, USA. 12 School of Geographical Sciences, Northeast Normal University, Changchun 130024, China. 13 Department of Geography, Portland State University, Portland, OR 97201, USA. 14 INIBIOMA, CONICET-Universidad Nacional del Comahue, Quintral 1250, Bariloche, 8400 Rio Negro, . 15 Clark University, Graduate School of Geography, Worcester, MA 01602, USA. 16 Fenner School of Environment and Society, The Australian National University, Canberra, ACT 2601, . 17 Graduate School of Environment and Information Sciences, Yokohama National University, Yokohama 240-8501, Japan. 18 Field Station Fabrikschleichach, Department of Animal Ecology and Tropical Biology, Biocenter, University of Würzburg, Glashüttenstraße 5, 96181 Rauhenebrach, Germany. 19 Bavarian Forest National Park, Freyunger Str. 2, 94481 Grafenau, Germany. 20 School of Environment, University of Auckland, Auckland 1142, New Zealand. 21 Department of , Policy and Management, University of California, Berkeley, CA 94720, USA. 22 Department of Integrative Biology, Birge Hall, University of Wisconsin–Madison, Madison, WI 53706, USA. 23 Department of Geography, University of Colorado, Boulder, CO 80309, USA. These authors contributed equally: Andreas Sommerfeld, Cornelius Senf. Correspondence and requests for materials should be addressed to A.S. (email: [email protected])

NATURE COMMUNICATIONS | (2018) 9:4355 | DOI: 10.1038/s41467-018-06788-9 | www..com/naturecommunications 1 ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-06788-9

atural disturbances are an essential component of forest local factors such as the topographic template of a landscape16 and ecosystems1. Yet, forest disturbance regimes are changing the influence of human land use17. A better understanding of N 2 in response to global . Hotter and pro- global disturbance patterns and their multi-scale drivers is also longed droughts3, exceptional bark beetle outbreaks4, and mega- crucial for improving the representation of disturbances in global fires5 have been increasingly reported in recent years, and are models18,19, and can have important implications for impacting forest across all forested continents2. policy decision making, e.g., in the context of climate change Changes in disturbance dynamics can have substantial impacts on mitigation20,21. services provided by , e.g., climate regulation6,7, New opportunities for global forest disturbance research arise provisioning of drinking water8,9, and protection from natural from recent advances in remote sensing. Increasingly available hazards10, as well as affect conservation of biological diversity11. remotely sensed datasets on forest disturbance22 offer high spatial Quantifying disturbance patterns (i.e., the size, shape, and pre- resolution and are globally consistent, enabling large-scale com- valence of disturbances in forest landscapes) and understanding parative efforts. However, while the global mapping of forest their drivers is thus a key challenge for ecological research. disturbances is now feasible23,24, attributing disturbance agents While the potential drivers of the ongoing disturbance change from remote sensing data and distinguishing between natural and are global, the responses to these drivers vary considerably at the anthropogenic disturbances remains challenging25. Furthermore, local scale. Insights from well-studied systems such as Yellowstone ecological context information such as the prevailing tree species National Park in North America1, the Bohemian Forest ecosystem composition cannot usually be gleaned from space, underlining in Europe12, and the O’Shannassy water catchment in Australia13 the importance of terrestrial information and local ecosystem have provided important insights into the complex interactions understanding for a meaningful interpretation of remotely sensed between climate variability, disturbances, and forest development. disturbance information. Integrating remote sensing analyses While an in-depth understanding of disturbance dynamics exists with in-depth knowledge on selected ecosystems across the globe for a growing number of landscapes (i.e., contiguous land areas of can provide new insights by combining the consistent synoptic roughly between 103 and 106 ha) around the globe, their responses view of satellite analysis with the expertise and insights gained to global drivers have not consistently been compared to date. from decades of local forest disturbance research. Questions such as whether recent bark beetle outbreaks in North Our objective was to analyze patterns and drivers of recent America differ from those in with regard to their climate disturbances across temperate forests at the global scale. We sensitivity, or whether recent fires in Australia created similar compiled a global network of 50 protected forest landscapes each patterns as those in the Americas remain largely unexplored. with > 2000 ha contiguous forest area (Fig. 1) for which in-depth Comparing the variation in disturbance patterns and their rela- local systems knowledge exists. We jointly analyzed severe canopy tionship to climate variability among landscapes at subcontinental disturbances (i.e., complete mortality of all taller 5 m within to global scales14,15 has the potential to elucidate whether recent a 30 m grid ) in these landscapes using Landsat-derived dis- disturbance episodes were consistently driven by climate across turbance maps for 2001–201424. Focusing our network of land- , or whether climate sensitivities differ between systems. scapes on protected areas and their surroundings allowed us to Furthermore, such a comparison can shed light on how regional- isolate patterns of natural disturbances (inside protected areas) to continental-scale drivers such as climate variability interact with from those of areas where natural and human disturbances

17 45 a 30 11 25 16 22 39 1 12 5 15 36 21 49 38 6 37 50 24 9 20 31 13 40 29 32 48 b 3000

3 47 10 7 35 27 43 4 1 18 19 2000 33 42 8 41 19 26 17 34 26 8 28 43 38 31 21 46 37 46 34 39 10 4 2 30 29 36 6 14 23 32 24 9 44 5 23 42 1000 47 28 48 40 11 45 22 2 18 27 12 50 15 16 20 14 3 41 Mean annual [mm] Mean annual 7 Temperate biome 13 49 25 33

0 44 35

–10 0 10 20 Mean annual temperature [°C]

Fig. 1 A network of 50 protected landscapes to understand global patterns and drivers of temperate forest disturbances. a The geographic location of the landscapes, and b their location in climate space. The area of the temperate biome is indicated in green47. See Supplementary Table 1 for more detailed information on the individual landscapes. Note that the climatic envelope of the biome (green dots in b) is based on a sample of 10,000 4500 m×4500 m grid cells throughout the biome

2 NATURE COMMUNICATIONS | (2018) 9:4355 | DOI: 10.1038/s41467-018-06788-9 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-06788-9 ARTICLE

Table 1 Characteristics of disturbance clusters

Cluster Low Moderate High Number of landscapes 18 23 9 Total forest area [ha] 788,986 1,216,364 1,965,572 Mean annual temperature [°C] 6.5 (5.6–7.5) 5.3 (4.4–6.2) 3.7 (2.8–4.6) Mean annual precipitation [mm] 1393 (1241–1544) 1222 (1071–1374) 1197 (1046–1349) Mean percent of forest area disturbed 2001–2014 0.31 (0.13–0.48) 4.61 (0.44–8.79) 21.50 (13.86–29.18) [%] Edge density [m/ha] 2.87 (1.24–4.50) 21.69 (2.80–40.58) 43.22 (25.53–60.91) Area-weighted mean patch size [ha] 0.66 (0.46–0.85) 24.22 (6.96–41.47) 4451.04 (365.24–8536.84) Area-weighted mean perimeter-area-ratio [m/ha] 960.09 (905.26–1014.92) 617.28 (560.74–673.82) 215.31 (150.15–280.74)

Characteristics of three global clusters of disturbance activity, determined based on satellite-derived disturbance metrics using Gaussian finite mixture models. Values in parentheses indicate the 95% confidence interval

interact (outside protected areas). We concentrated our analysis topographic complexity, and the spatio-temporal dynamics of on a single biome as we were particularly interested in within- forest disturbances. If recent disturbance episodes are responding biome variation in disturbance patterns and drivers, rather than consistently to climatic and topographic drivers, we would expect the comparatively well-studied between-biome differences. We to find a non-random signal in a regression analysis across our set selected temperate forests as the target of our study as they are of globally distributed landscapes (H3). Alternatively, if responses affected by a wide variety of disturbance agents, frequently con- vary among landscapes and cancel each other out at the global tain both angiosperm and gymnosperm species (i.e., high trait scale, the regression coefficients for these drivers would not differ variability), and are represented in both the northern and significantly from zero. southern hemisphere, spanning an extensive gradient in envir- onmental conditions. Results We hypothesized that differences in disturbance agents and Patterns of recent natural disturbances. Disturbance dynamics tree species are the main determinants of among-landscape var- between 2001 and 2014 varied strongly across the temperate fi iation in disturbance patterns across the globe (H1). Speci cally, forest biome. Unsupervised cluster analysis identified three dis- fi we expected that areas predominately affected by re as well as tinct groups of landscapes based on their recent disturbance landscapes dominated by tree species with high general suscept- dynamics (Table 1; Supplementary Figure 1), which we in the ibility (i.e., traits such as high maximum tree height and low following refer to as low, moderate, and high disturbance activity wood density) are most affected by disturbances. Our alternative clusters. Each cluster comprises a group of landscapes of similar hypothesis was that spatial proximity of landscapes is a good characteristics with regard to the size and shape of disturbance indicator of similarities in disturbance patterns, with global dif- patches, the percentage of a landscape disturbed during the study ferences in disturbances mainly explained by geographical loca- period, and the average amount of edges created by these dis- tion (e.g., on different continents or hemispheres). To test these turbances (Supplementary Figure 2). Approximately one-third of hypotheses, we calculated four disturbance indicators, whereof the landscapes studied (representing 19.9% of the forest area) fell two are landscape-level metrics (percent of landscape disturbed within the low disturbance activity cluster. This group was – 2001 2014, edge density) and two are patch-level metrics (area- characterized by small and complex disturbance patches weighted mean patch size, area-weighted mean perimeter-area- (Table 1), with disturbances on average affecting only 0.31% of ratio), and used cluster analysis to identify patterns among the landscape’s forest area between 2001 and 2014. Examples of landscapes indicating differences in recent disturbance activity landscapes with low disturbance activity are the Te Paparahi (with disturbance activity jointly referring to the prevalence, size, Conservation Area (New Zealand), Shiretoko (Japan), Feng Lin and shape of disturbances as characterized by our four focal (China), Five (USA), Hainich (Germany), and Hornopirén indicators). Furthermore, we hypothesized that the patterns of (Chile). The majority of the landscapes (23 landscapes, repre- natural disturbances (i.e., those observed in protected areas with senting 30.6% of the forest area studied) fell within the moderate fl only minimal direct human in uence) differ from the combined disturbance activity cluster. This group was characterized by a natural and human disturbances outside protected areas (H2). roughly 30 times larger area-weighted mean patch size than the We expected disturbances in protected areas to be generally landscapes in the low disturbance activity cluster (Table 1). smaller and more complex in shape (i.e., higher perimeter-area- Disturbance patches in the moderate cluster were less complex ratio) compared to those outside of protected areas. This (i.e., had a lower area-weighted mean perimeter-area-ratio) but hypothesis is based on the insights that disturbances outside affected a larger forest area (on average 4.61% of the landscape) protected areas are the result of both natural and human dis- between 2001 and 2014. Examples of landscapes in this group are turbances (which can amplify each other, e.g., when strong winds the Bavarian Forest (Germany), Baxter State Park (USA), Los uproot edge trees of freshly created clear-cuts), and that Alerces (Argentina), and Nelson National Park (New management-related biotic homogenization has the potential to Zealand). Although only 9 of the 50 landscapes analyzed fell increase forest susceptibility to disturbances relative to natural within the high disturbance activity cluster, they accounted for 26–28 ecosystem development . The alternative hypothesis was that 49.5% of the total forest area under study. Area-weighted mean high recent levels of natural disturbance activity supersede the patch size was two orders of magnitude larger than in the signal of human land use, with similar disturbance patterns inside moderate disturbance activity cluster, and disturbance patches and outside protected areas. Finally, based on local and regional were considerably less complex (Table 1). On average, almost one 29 studies highlighting the importance of climate variability and quarter of the landscape’s forested area was affected by dis- 30 landscape structure for disturbance dynamics, we tested for a turbances between 2001 and 2014 in the high disturbance activity consistent global relationship among climate variability, relative cluster, resulting in the highest edge density among all three

NATURE COMMUNICATIONS | (2018) 9:4355 | DOI: 10.1038/s41467-018-06788-9 | www.nature.com/naturecommunications 3 ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-06788-9

a ● 0.1● 0.2 ● 0.3 0.4 b ● 0.1 ● 0.2 ● 0.3

Windthrow ● ● Nothofagus ●

Picea ● ● ● Wildfires ● ●

Abies ● ● ● Bark beetles ● ● Fagus ● ● Drought ● ● ● Pseudotsuga ● Pathogens ● ● Eucalyptus ●

Defoliators ● Pinus ● ●

Snow & ice ● ● Acer ●

Volcanism ● Tsuga ●

Low Moderate High Low Moderate High

c 100 ● ● p = 0.57 100 p = 0.19 1.0 p = 0.02 p = 0.04 ] p = 0.53 p = 0.04 3 p = 0.02 p = 0.13 p = 0.53 p = 0.19 p = 0.76 80 p = 0.79 100 ● ● ● 75 0.8 ● 60 50 50 0.6 [%] Conifers Dominance [%] 40

25 density [mg/mm Wood

Maximum tree height [m] Maximum ● 0 ● 0.4 ●

Low High Low rate High Low High Low High Moderate Mode Moderate Moderate

Fig. 2 Distribution of disturbance agents, tree genera, and tree species traits across three global clusters of disturbance activity (cf. Table 1). Bubbles are scaled relative to the occurrence of the two most important a disturbance agents and b tree genera within each cluster. c Dominance [%] indicates the share of the single most prevalent tree species on the overall tree species composition, while conifers [%] indicates the respective share of all species. Maximum tree height and wood density indicate a weighted trait distribution across landscapes in the respective disturbance activity clusters. Boxplots denote the median (center line) and interquartile range (box), with whiskers extending to three times the interquartile range and points indicating values outside this range. Test statistics and p-values are based on approximate Kruskal–Wallis tests with 9999 permutations. For further information on statistical analyses see Supplementary Table 2 clusters. Examples for landscapes in this group are Yellowstone from this group of landscapes. In landscapes experiencing (USA), Puyehue (Chile), and O’Shannassy (Australia). moderate disturbance activity in 2001–2014, fire and bark beetle The clustering based on the four landscape metrics considered outbreaks were more prevalent compared to the low cluster. High here did not reveal strong geographical patterns (Supplementary disturbance activity was predominately associated with wildfire, Figure 3a). Landscapes from at least three continents were present with bark beetle outbreaks and drought also being important in each cluster, and all three disturbance activity clusters were agents in highly disturbed landscapes. represented in both the southern and northern hemispheres. Dominant tree genera (i.e., the genera with the highest Furthermore, no clear pattern emerged when comparing the proportion of basal area) differed among disturbance activity clusters in climate space (Supplementary Figure 3b), although a clusters (χ2 = 69.09, p < 0.001). Low disturbance activity land- tendency of cooler landscapes experiencing higher disturbance scapes were largely dominated by broadleaved trees from the activity could be detected (Table 1). genera Nothofagus, Fagus, and Acer, i.e., species that have a Disturbance activity clusters were associated with different relatively low maximum attainable height but higher wood major disturbance agents and tree genera (Fig. 2a, b, Supple- density (Fig. 2c). The moderate disturbance activity cluster was mentary Figure 4). Agents differed significantly among dis- characterized by both broadleaved and coniferous tree species, yet turbance activity clusters (χ2 = 37.64, p < 0.01). Landscapes in the their average trait characteristics largely resembled those of the low disturbance activity cluster were frequently affected by low disturbance activity landscapes. High disturbance activity multiple disturbance agents, with windthrow being the most landscapes in the northern hemisphere were dominated by the prevalent agent. Major bark beetle outbreaks were largely absent genera Picea, Abies, Pseudotsuga, and Pinus, which are

4 NATURE COMMUNICATIONS | (2018) 9:4355 | DOI: 10.1038/s41467-018-06788-9 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-06788-9 ARTICLE

a b p < 0.01 p < 0.01 p = 1 p < 0.01 p < 0.01 p = 0.48

10,000 ● 0.12

●

Outside 100 0.08

● Inside patch size [ha]

Area−weighted mean Area−weighted mean 0.04 1 perimeter−area−ratio [m/sqm]

0.00 Low Moderate High Low Moderate High

Fig. 3 Comparison of disturbance patterns inside and outside protected areas. a Area-weighted mean patch size and b area-weighted mean perimeter- area-ratio are compared for areas inside and outside protected areas for three global clusters of disturbance activity (cf. Table 1). Boxplots denote the median (center line) and interquartile range (box), with whiskers extending to three times the interquartile range and points indicating values outside this range. Test statistics and p-values are based on approximate Kruskal–Wallis tests with 9999 permutations characterized by a higher maximum attainable tree height and negative effect of temperature anomaly (GLMM; β = −0.20, std. lower wood density (see Supplementary Table 2 for test statistics error = 0.07, p = 0.01; see Supplementary Table 4). That is, on trait differences between clusters). Conversely, the high warmer than average temperatures decreased the probability of disturbance activity landscapes located in the southern hemi- disturbance in the following years (Fig. 4a). This effect of tem- sphere were mainly characterized by Nothofagus. The share of the perature was independent of variation in precipitation. Pre- single most dominant tree species on the overall species cipitation anomalies had a significant positive direct effect on composition did not differ between landscapes in the low and disturbance probability (GLMM; β = −0.33, std. error = 0.07, p < moderate clusters. In high disturbance activity landscapes, 0.01; Supplementary Table 4), with wetter conditions increasing however, the most important tree species was less dominant disturbance probability in the low disturbance activity cluster compared to landscapes in low and moderate clusters (Fig. 2c). (Fig. 4a). Conversely, for landscapes in the high disturbance activity cluster, we found a significant positive effect of tem- perature (GLMM; β = 0.59, std. error = 0.01, p < 0.01; Supple- Disturbance differences inside and outside protected areas. mentary Table 4), with a higher disturbance probability in years Disturbance patches inside protected areas—almost exclusively with above average temperature (Fig. 4c). This effect was further influenced by natural disturbance agents (but see Supplementary modulated by precipitation, and was strongly amplified if warm Table 6)—were smaller and more complex than disturbance years coincided with drier than average conditions (GLMM; β = patches in surrounding areas affected by both human and natural −0.43, std. error < 0.01, p < 0.01; Supplementary Table 4). While disturbances in the low and moderate disturbance activity clusters landscapes in the moderate disturbance activity cluster showed an (Fig. 3, Supplementary Table 3). For landscapes with high dis- overall negative effect of temperature on disturbance probability turbance activity, however, the distribution of patch sizes and β = − = fi (GLMM; 0.30, std. error 0.02, p < 0.01; Supplementary perimeter-area-ratios did not differ signi cantly between pro- Table 4), the effect was also significantly influenced by pre- tected areas and their surroundings. Hence, there is a higher cipitation (GLMM; β = −0.17, std. error = 0.03, p < 0.01; Sup- similarity between disturbances in protected and unprotected plementary Table 4). In addition, disturbance probability systems in areas that experienced high disturbance activity increased following years with above average temperature and recently. Furthermore, patch size and patch complexity differed below average precipitation (Fig. 4b). more strongly among disturbance activity clusters in protected Relative topographic complexity significantly affected spatial systems compared to unprotected systems. disturbance patterns, with different effects across disturbance activity clusters (Supplementary Table 4). In the predominately Drivers of spatio-temporal disturbance dynamics. Inter-annual wind-influenced low disturbance activity landscapes, disturbance climate variability was an important driver of temporal dis- probability increased by ~7.5% with an increase in topographic turbance dynamics in all three disturbance activity clusters. The complexity by one standard deviation (GLMM; β = 0.30, std. full model (including both temperature/precipitation anomalies error = 0.04, p < 0.01; Supplementary Table 4). In contrast, and topographic complexity as predictors) was more strongly disturbance probability decreased by ~1.25% (GLMM; β = supported by the data than the spatial-only model (including only −0.05, std. error = 0.01, p < 0.01; Supplementary Table 4) and topographic complexity) and the Null model (likelihood-ratio 2.5% (GLMM; β = −0.10, std. error < 0.01, p < 0.01; Supplemen- tests; all p-values < 0.01; see Supplementary Table 4). However, tary Table 4) with the same amount of change in topographic the direction and strength of effects, as well as the lag time of complexity in the mainly beetle-driven and fire-driven landscapes climate effects, varied among clusters (Supplementary Figure 5; of the moderate and high disturbance activity clusters. Fig. 4). A 2-year and 3-year lag was most strongly supported by the data for the low and moderate disturbance activity clusters, respectively. Conversely, climate variability had an immediate Discussion influence on disturbances (i.e., zero lag) in the high disturbance We present a quantitative analysis of recent disturbance dynamics activity cluster (Supplementary Figure 5). For landscapes with low across the temperate forest biome. Variation in recent forest disturbance activity, we found a significant but moderately disturbance activity across the globe was considerable, spanning a

NATURE COMMUNICATIONS | (2018) 9:4355 | DOI: 10.1038/s41467-018-06788-9 | www.nature.com/naturecommunications 5 ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-06788-9

abcLow Moderate High 0.012 0.060 0.800

0.009 0.600 0.040 Precipitation anomaly 0.006 0.400 −2 0.020 (disturbance)

P 0.003 0.200 0

2 0.000 0.000 0.000 −2 −1 0 1 2 −2 −1 0 1 2 −2 −1 0 1 2 Temperature anomaly

Fig. 4 Predicted response of disturbance probability to temperature anomaly, modulated by precipitation anomaly. a–c The climate sensitivity of disturbances separately for three global clusters of disturbance activity (cf. Table 1). Anomaly values are units of standard deviation with zero indicating the long-term mean. Y-axes are scaled differently across the three panels for clarity of presentation. Prediction uncertainty was estimated from 9999 model simulations, with the lower and upper limit representing the 2.5 and 97.5% quantile of all simulations. We note that prediction intervals include both parameter and model uncertainty, and overlapping prediction intervals can occur despite significant differences in parameter values. For parameter estimates and standard errors, see Supplementary Table 4. The number of experimental replicates equals the number of study sites per cluster (Low: 18, Moderate: 23, High: 9) large gradient of patch sizes and landscape area affected by dis- areas, supporting our expectation (H2). In landscapes affected by turbance. Our results highlight that while extensive disturbances large-scale disturbances, however, the patterns in protected areas such as massive bark beetle outbreaks or severe large-scale fires and their surroundings were similar. This suggests that large have garnered considerable attention from researchers and the natural disturbances can override the effect of human land use public recently, many temperate forests are dominated by small- and dominate landscape patterns in forest ecosystems. As these scale disturbance events. This finding underscores the importance events have been found to be particularly climate sensitive of a consistent quantification and analysis of disturbance (Fig. 4c), future conditions could produce more coarse-grained dynamics across systems. The main disturbance agent affecting a landscape patterns (i.e., landscapes characterized by larger patch system was more indicative of within-biome variation in dis- sizes) in temperate forests36. turbance activity than geographical proximity of landscapes or An important caveat associated with our analysis is that we their location on the same or hemisphere (H1). Spe- could not consider vegetation structure and disturbance history as cifically, wind was an important agent responsible for small-scale potential covariates in our analyses, resulting from the lack of a disturbances in temperate forests31. Wildfires and bark beetle globally consistent data set on these variables. It therefore remains outbreaks, on the other hand, were the two most prominent unclear whether the differences between protected and unpro- agents associated with large and severe disturbances in recent tected areas arise from a higher disturbance susceptibility of the years. However, the fact that both wind and wildfires occurred in latter systems, or whether they reflect management activities such all three clusters of disturbance activity highlights the consider- as in areas outside protected landscapes. Disturbance able within-biome variability even within disturbance regimes history and ecological legacies can exert an important influence characterized by the same agent. on current disturbance activity37,38. This is important as most of Tree species composition was related to global differences in our study landscapes have been protected only for a few decades, recent disturbances, with Northern Hemisphere temperate forests and legacies from former land use might still persist. Further- dominated by conifers experiencing elevated disturbance activity. more, although our 50 protected landscapes cover the climatic This pattern can partly be explained by the general life-history envelope of temperate forests well (Fig. 1b), they are not neces- strategy of conifers and their extensive coevolution with dis- sarily representative of the full range of ecological conditions of turbances32, as many conifers are well adapted to either tolerate the entire temperate forest biome. Selecting areas with a long disturbances or swiftly recolonize disturbed areas. An unexpected history of forest dynamics research enabled us to consistently finding was that the dominance of the single most prevalent tree compare patterns and responses across locally well-researched species was lower in landscapes with high disturbance activity systems, but might also be a source of bias that should be con- compared to those with low and moderate disturbance activity. sidered in interpreting our results. The finding of similar dis- This contrasts with previous suggestions that disturbance risk turbance patterns inside and outside of protected areas in the increases if landscapes are dominated by a small number of tree high disturbance activity cluster may, for instance, partly result species33. However, disturbances themselves can have a positive from the generally remote location of these particular landscapes, effect on tree species diversity34,35. Consequently, higher evenness with limited human activity in their surroundings. Future ana- in systems strongly affected by disturbances—as found here— lyses explicitly contrasting disturbances inside and outside pro- could be a consequence of disturbances, rather than being cau- tected areas could help to better understand the effect of human sally related to them. activity on disturbances, and quantify the impact of anthro- In large parts of the temperate forest biome, human dis- pogenically altered disturbance regimes on biodiversity25,39. turbances dominate the landscape. Consequently, the disturbance Furthermore, we here focused on severe canopy disturbances with patterns of unprotected systems differed from the natural complete canopy mortality of all trees taller 5 m within a 30 × 30 disturbance regime observed in protected areas. In the majority m grid cell24. Consequently, low severity disturbances and of human-dominated landscapes outside protected areas, dis- understory tree mortality were not considered, potentially leading turbance patches were larger and less complex than in protected to an underestimation of forest disturbance activity. However, the

6 NATURE COMMUNICATIONS | (2018) 9:4355 | DOI: 10.1038/s41467-018-06788-9 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-06788-9 ARTICLE severe canopy disturbances examined here are generally well resilience and adaptive capacity of ecosystems to disturbances represented by the data set used: For example, Buma and Bar- remain important priorities of forest research and management. rett40 found an overall agreement of 91% when comparing the global disturbance data set used here to high-resolution imagery Methods in southern Alaska. Furthermore, Borelli et al.41 determined an A biome-wide network of protected forest landscapes. We compiled a network overall accuracy of 81% in an independent evaluation of the data of study landscapes distributed throughout the temperate forest biome as defined 47 across Europe. We are thus confident that the data set used here by Olson et al. . (see also Fig. 1). Selection criteria were that the landscapes are protected (i.e., IUCN Cat. I and Cat. II), and have a minimum of 2000 ha of is able to capture the variability in severe canopy disturbances contiguous forest area. Studying protected areas allowed us to largely control for across the temperate forest biome. anthropogenic disturbances and focus our main analyses on natural disturbances. A remaining limitation of our analysis is the relatively short We analyzed 50 landscapes distributed across 16 countries on five continents, duration of our study period. The currently available disturbance representing a forest area of 3.9 Mill. ha (median landscape size: 30,889 ha; see Supplementary Table 1 for details). The study landscapes cover a wide climatic time series from satellite data remain too short to characterize gradient of the temperate forest biome, with mean annual temperatures ranging 1 disturbance regimes satisfactorily, and preclude the assessment of from −0.3 °C to 14.8 °C, and mean annual precipitation sums between 517 mm temporal trends in disturbance7,42. Consequently, we focused only and 2315 mm (Supplementary Table 1 and Fig. 1b). on the effect of inter-annual climate variability rather than on long- term trends, using temperature and precipitation anomalies as Disturbance data and landscape pattern analysis. We acquired forest cover and 24 predictors. Future work should also consider the effect of climatic annual disturbance maps (2001–2014) from Hansen et al. (Version 1.2) at 30 m fi spatial resolution. A disturbance was defined as a severe canopy disturbance, extremes and intra-annual climate patterns for re ning our meaning the complete mortality of all trees taller 5 m within a pixel24. Only dis- 29 understanding of climate—disturbance relationships .Inaddition, turbance events that occurred between 2001 and 2014 were considered. To char- process-based simulation modeling43 could be employed to obtain acterize disturbance patterns, we calculated two landscape-level metrics and two a more dynamic and long-term perspective on global disturbance patch-level metrics for each study landscape, using an eight-neighbor rule for defining adjacency and considering disturbances throughout the entire study regimes and their responses to a changing climate. period. The landscape-level metrics were: (i) percent of landscape disturbed Here we provide evidence that high recent disturbance activity 2001–2014, and (ii) edge density of all disturbed patches within the forest area of a in temperate forest ecosystems across the globe was strongly landscape; with the patch-level metrics being (iii) area-weighted mean patch size, related to the joint occurrence of warmer and drier than average and (iv) area-weighted mean perimeter-area-ratio. To identify differences and fi similarities in recent disturbance patterns across the temperate forest biome, we climate conditions (H3). These global scale ndings are in general fi 48 4,14,42,44,45 used Gaussian nite mixture models, as implemented in the R package mclust agreement with local studies , particularly considering (version 5.4). Gaussian finite mixture models are an approach for unsupervised that our high disturbance activity cluster was dominated by clustering, used here to identify groups of landscapes with similar disturbance wildfires, bark beetle outbreaks, and drought. Our results there- patterns. The optimal number of cluster centers was determined by maximizing the fore suggest that a warming climate could facilitate large-scale Bayesian Information Criteria (BIC). Robust statistics for characterizing each 2,46 cluster were derived by using parametric bootstrapping (9999 replications). Sub- disturbances in temperate forest ecosystems in the future . Our sequently, the clusters were characterized with regard to their main disturbance findings also show that climate sensitivity can, to some degree, be agents and tree genera (i.e., the two most important agents and genera for a buffered by heterogeneous topography, which impedes the spread landscape during the period 2001–2014; see below for details). To describe of disturbances and/ or increases the complexity of disturbed potential functional differences between the clusters, we included traits in our 25 analysis. After an initial screening and analysis of multicollinearity, we focused on patches . Interestingly, our analysis suggests that both the effect two complementary plant traits corresponding to disturbance resistance and sus- of climate and the effect of topography reversed for landscapes ceptibility, i.e., maximum potential tree height and mean wood density, extracted characterized by low disturbance activity, compared to those with from the TRY database49. Plant height directly increases susceptibility to wind high disturbance activity. For the former, which are pre- disturbance and is also a proxy for accumulation potential, which is related to fuel load in the context of disturbances by wildfire. Wood density is dominately driven by wind disturbance, cooler and wetter con- positively correlated with the ability of trees to resist physical forces such as wind ditions as well as higher relative topographic complexity and drought. For each landscape, weighted trait means based on tree species shares increased disturbance probability31, which is consistent with were calculated. Differences in agents and tree genera among clusters were tested decreased tree anchorage (soil wetness) and increased wind sus- using approximate Pearson χ2 tests of homogeneity with 9999 permutations. Dif- ceptibility (exposed ridges, funnel effects) under such conditions. ferences in traits among clusters were tested using two-tailed pairwise approximate Kruskal–Wallis tests with 9999 permutations, applying false discovery rate cor- However, the signals detected for low disturbance activity areas rection. All test procedures were used as implemented in the coin50 package were generally weak, underlining the highly stochastic nature of (version 1.2-1) in R51. small-scale mortality events in forest ecosystems. We conclude by emphasizing the importance of protected areas Expert-based information on local ecological context. Remote sensing provides for understanding changes in forest landscapes in the absence of a consistent estimate of disturbances across the biome, yet the ecological context of direct human influences. Furthermore, our work highlights the these disturbances, such as the dominant disturbance agents and the tree species affected, cannot be inferred from space. We thus consulted local experts for all 50 importance of consistent global information for characterizing landscapes, collecting ecological context information via a questionnaire (see patterns and identifying drivers of important ecological processes Supplementary Table 5). The questionnaire included four questions, two of which such as disturbances. Quantitative baselines acknowledging the were multiple choice questions with the opportunity to add additional answers. substantial spatio-temporal variability in ecosystems are needed to They focused on determining the dominant tree species, the main disturbance agents affecting a landscape between 2001 and 2014, and the impact of dis- identify, monitor, and attribute changes in ecological processes. turbances on particular tree species. Values of tree species dominance in a given The analyses presented here are an important step towards such an landscape were estimated as basal area shares. Local experts were identified via improved quantitative characterization of forest disturbances at the their publication record on the topic of forest dynamics for the selected areas. All consulted experts also contributed to the analysis and interpretation of the data, global scale, combining large-scale remote sensing data with eco- fi logical context information from local experts. An improved and are identi ed individually in Supplementary Table 1. quantitative characterization of forest disturbances at the global Difference inside and outside protected areas. After focusing on natural dis- scale can, for instance, inform the development and application of turbance dynamics in protected areas in the previous analysis steps, we asked how global vegetation models, which largely ignore the impacts of protected areas differed from the disturbance dynamics in the unprotected systems disturbances to date, or only consider a highly simplified repre- surrounding our study landscapes. To compare spatial disturbance patterns inside sentation of disturbance processes18,19. An improved consideration vs. outside protected areas, we extracted all forest disturbances in a buffer sur- rounding the protected landscapes. The buffer size was selected proportional to the of disturbance processes in future projections is important as our landscape size, and was set to the diagonal of the minimum bounding rectangle of results highlight the considerable sensitivity of disturbances to the each study landscape. We compared area-weighted mean patch size and area- ongoing changes in the climate system. We conclude that the weighted mean perimeter-area-ratio between strata (i.e., inside vs. outside

NATURE COMMUNICATIONS | (2018) 9:4355 | DOI: 10.1038/s41467-018-06788-9 | www.nature.com/naturecommunications 7 ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-06788-9 protected areas) using boxplots, and tested differences using two-tailed approx- 3. Millar, C. I. & Stephenson, N. L. Temperate forest health in an era of emerging imate Kruskal–Wallis tests with 9999 permutations. megadisturbance. Science 349, 823–826 (2015). 4. Raffa, K. F. et al. Cross-scale drivers of natural disturbances prone to fi Drivers of spatio-temporal disturbance dynamics. We used generalized linear anthropogenic ampli cation: the dynamics of bark beetle eruptions. Bioscience – mixed effects models52 (GLMMs) to test the influence of relative topographic 58, 501 517 (2008). fi complexity and climatic variability on spatial and temporal disturbance dynamics. 5. Stephens, S. L. et al. Temperate and boreal forest mega- res: characteristics – Analyses were run separately for each disturbance activity cluster, modeling the and challenges. Front. Ecol. Environ. 12, 115 122 (2014). annual probability of disturbance at the pixel level. As response variable we used 6. Bright, B. C., Hicke, J. A. & Meddens, A. J. H. Effects of bark beetle-caused tree annual binary disturbance maps indicating whether a pixel was disturbed in a given mortality on biogeochemical and biogeophysical MODIS products. J. Geophys. year or not. Consequently, the GLMMs were specified with a binomial error dis- Res. Biogeosciences 118, 974–982 (2013). tribution and a logit link-function. As measure of topographic complexity we used 7. Seidl, R., Schelhaas, M. J., Rammer, W. & Verkerk, P. J. Increasing forest the topographic ruggedness index53 (TRI), which was calculated from a 30 m disturbances in Europe and their impact on carbon storage. Nat. Clim. Chang. digital elevation model obtained from the Shuttle Radar Topographic Mission 4, 806–810 (2014). (SRTM). We calculated the TRI using a window size of 7 × 7 pixels, depicting the 8. Rajan, S., Firdaus, N. N. M., Appukutty, M. & Ramasamy, K. Effects of climate topographic complexity within a radius of ~100 m around a focal pixel. TRI values changes on dissolved heavy metal concentrations among recreational park were subsequently scaled to zero mean and a standard deviation of one for each tributaries in Pahang, Malaysia. Biomed. Res. 23,23–30 (2012). landscape, with negative values indicating a lower than average relative topographic 9. Mikkelson, K. M., Dickenson, E. R. V., Maxwell, R. M., Mccray, J. E. & Sharp, complexity, and positive values indicating a higher than average relative topo- J. O. Water-quality impacts from climate-induced forest die-off. Nat. Clim. graphic complexity. As a measure of climate variability we obtained time series of Chang. 3, 218–222 (2013). mean annual temperature and annual precipitation sum from FetchClimate54, 10. Wohlgemuth, T., Schwitter, R., Bebi, P., Sutter, F. & Brang, P. Post-windthrow which are based on daily climate data from the NCEP/NCAR Reanalysis 1 data- management in protection forests of the Swiss Alps. Eur. J. For. Res. 136, base. We calculated climate anomalies by scaling the time series to zero mean and a 1029–1040 (2017). standard deviation of one for each landscape, with negative values indicating 11. Beudert, B. et al. Bark beetles increase while maintaining drinking colder/ dryer than average years, and positive values indicating hotter/ wetter than water quality. Conserv. Lett. 8, 272–281 (2015). average years. As previous studies suggest variable lag times between climate 12. Seidl, R. et al. Small beetle, large-scale drivers: how regional and landscape 12,16 anomalies and disturbance signals determined from remote sensing , we tested factors affect outbreaks of the European spruce bark beetle. J. Appl. Ecol. 53, variable lags ranging from 0 years (i.e., relating the climate anomaly of the current 530–540 (2016). year to the disturbance in the current year) to 3 years (i.e., relating the climate 13. Lindenmayer, D. B., Hobbs, R. J., Likens, G. E., Krebs, C. J. & Banks, S. C. anomaly 3 years prior to the disturbance in the current year). The lag best sup- fi ’ Newly discovered landscape traps produce regime shifts in wet forests. Proc. ported by the data within each cluster was identi ed using Akaike s Information Natl Acad. Sci. USA 108, 15887–15891 (2011). Criterion (AIC)16,55. Furthermore, we allowed for an interaction term between 14. Allen, C. D. et al. A global overview of drought and heat-induced tree precipitation and temperature to account for potential modulating effects between mortality reveals emerging climate change risks for forests. For. Ecol. Manag. these two variables56. Using GLMMs enabled us to incorporate factors on different 259, 660–684 (2010). hierarchical levels into a combined modeling framework. In particular, TRI values 15. Waring, R. H., Coops, N. C. & Running, S. W. Predicting satellite-derived were available at the pixel level, but remain constant across years. In contrast, the patterns of large-scale disturbances in forests of the Pacific Northwest temperature and precipitation anomalies varied among years, but did not differ – spatially in a study landscape. Hence, we used relative topographic complexity to in response to recent climatic variation. Remote Sens. Environ. 115, 3554 3566 explain the spatial variation in disturbance probability, while climate variability was (2011). related to temporal variation of disturbances in our model. In addition, the GLMM 16. Senf, C. & Seidl, R. Natural disturbances are spatially diverse but temporally framework allowed us to account for random variation in the model intercept synchronized across temperate forest landscapes in Europe. Glob. Chang. Biol. – among study landscapes within a cluster. As sample sizes were very large (several 24, 1201 1211 (2018). millions of 30 m pixels), we randomly sampled 10% of the pixels per landscape. As 17. Spies, T. A. et al. Using an agent-based model to examine forest management fi disturbances were rare in many landscapes, we employed a case-control sampling outcomes in a re-prone landscape in Oregon, USA. Ecol. Soc. 22,25 design57, that is we randomly down-sampled the absence class (no disturbance) to (2017). approximately the same size of the presence class (disturbance). This design has the 18. Kautz, M., Anthoni, P., Meddens, A. J. H., Pugh, T. A. M. & Arneth, A. advantage that model estimates are unbiased (with the exception of the intercept), Simulating the recent impacts of multiple biotic disturbances on forest carbon and the intercept can easily be corrected using the known true proportion of cycling across the United States. Glob. Chang. Biol.. https://doi.org/10.1111/ disturbance presence/absence in the . Finally, we compared three can- gcb.13974 (2017). didate models per cluster: a full model containing predictors of spatial (TRI) and 19. Bonan, G. B. & Doney, S. C. Climate, ecosystems, and planetary futures: The temporal (precipitation and temperature anomalies as well as their interaction) challenge to predict life in system models. Science 359, eaam8328 variation, a model containing TRI only (spatial-only model), and a null model (2018). (containing only an intercept while maintaining the random effects structure of the 20. Kurz, W. A. et al. Mountain pine beetle and forest carbon feedback to climate GLMMs). Model comparison was done using the AIC and log- likelihood tests. For change. Nature 452, 987–990 (2008). the model most strongly supported by the data we created response curves by 21. Thom, D., Rammer, W. & Seidl, R. The impact of future forest dynamics on drawing 9999 random simulation from the model (i.e., accounting for parameter climate: interactive effects of changing vegetation and disturbance regimes. and model uncertainty), as suggested by Gelman and Hill58. All models and cal- Ecol. Monogr. 87, 665–684 (2017). culations were implemented using the lme4 package59 (version 1.1-14) in R. 22. Wulder, M. A. & Coops, N. C. Satellites: make Earth observations open access. Nature 513,30–31 (2014). 23. Kuemmerle, T. et al. Challenges and opportunities in mapping land use Data availability – Data on forest disturbances were derived from the global forest change data set24, intensity globally. Curr. Opin. Environ. Sustain. 5, 484 493 (2013). available at https://earthenginepartners.appspot.com/science-2013-global-forest. 24. Hansen, M. C. et al. High-resolution global maps of 21st-century forest cover – Data on tree species traits were derived from the plant trait database TRY49, change. Science 342, 850 853 (2013). fl available at https://www.try-db.org/TryWeb/Home.php (doi: 10.17871/TRY.3). 25. Senf, C., P ugmacher, D., Hostert, P. & Seidl, R. Using Landsat time series for Landscape-level data on ecological context variables was derived by means of a characterizing forest disturbance dynamics in the coupled human and natural – questionnaire (see Supplementary Table 5), and is published in full in systems of Central Europe. ISPRS J. Photogramm. Remote Sens. 130, 453 463 Supplementary Table 1. (2017). 26. Mori, A. S., Isbell, F. & Seidl, R. β-diversity, assembly, and ecosystem functioning. Trends Ecol. Evol. 0,1–16 (2018). Received: 28 March 2018 Accepted: 28 September 2018 27. Lindenmayer, D., Thorn, S. & Banks, S. Please do not disturb ecosystems further. Nat. Ecol. Evol. 1, 31 (2017). 28. Seidl, R., Donato, D. C., Raffa, K. F. & Turner, M. G. Spatial variability in tree regeneration after wildfire delays and dampens future bark beetle outbreaks. Proc. Natl Acad. Sci. USA 113, 13075–13080 (2016). 29. Neumann, M., Mues, V., Moreno, A., Hasenauer, H. & Seidl, R. Climate References variability drives recent tree mortality in Europe. Glob. Chang. Biol. 23, 1. Turner, M. G. Disturbance and landscape dynamics in a changing . 4788–4797 (2017). Ecology 91, 2833–2849 (2010). 30. Collins, B. M. & Stephens, S. L. Stand-replacing patches within a ‘mixed 2. Seidl, R. et al. Forest disturbances under climate change. Nat. Clim. Chang. 7, severity’ fire regime: quantitative characterization using recent fires in a long- 395–402 (2017). established natural fire area. Landsc. Ecol. 25, 927–939 (2010).

8 NATURE COMMUNICATIONS | (2018) 9:4355 | DOI: 10.1038/s41467-018-06788-9 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-06788-9 ARTICLE

31. Mitchell, S. J. Wind as a natural disturbance agent in forests: a synthesis. Acknowledgements Forestry 86, 147–157 (2013). A.S. and R.S. acknowledge support from the Austrian Science Fund (FWF) through 32. Loehle, C. Strategy space and the disturbance spectrum: a life‐history model START grant Y895-B25. C.S. acknowledges funding from the German Academic for tree species coexistence. Am. Nat. 156,14–33 (2000). Exchange Service (DAAD) with funds from the German Federal Ministry of Education 33. Perry, D. A. in Forest Ecosystems (eds Perry, D. A., Oren, R., Hart, S.C.) (Johns and Research (BMBF) and the People Programme (Marie Curie Actions) of the European Hopkins University Press, Baltimore, 2008) . Union’s Seventh Framework Programme (FP7/2007–2013) under REA grant agreement 34. Thom, D. & Seidl, R. Natural disturbance impacts on ecosystem services and Nr. 605728 (P.R.I.M.E.—Postdoctoral Researchers International Mobility Experience). T. biodiversity in temperate and boreal forests. Biol. Rev. Camb. Philos. Soc. 91, D. acknowledges funding from the Fonds institutionnel de recherche de l’Universitédu 760–781 (2016). Quebeć en Abitibi-Temiscamingue,́ the Natural Sciences and Engineering Research 35. Silva Pedro, M., Rammer, W. & Seidl, R. A disturbance-induced increase in tree Council of Canada (NSERC), Tembec, and EACOM Timber Corporation. Á.G.G. was species diversity facilitates forest . Landsc. Ecol. 31, 989–1004 (2016). supported by FONDECYT 11150835. S.J.H. and T.T.V. acknowledge NSF Award 36. Hansen, W. D., Braziunas, K. H., Rammer, W., Seidl, R. & Turner, M. G. It 1262687. A.H. was partially supported by NSF (award #1738104). D.K. acknowledges takes a few to tango: changing climate and fire regimes can cause regeneration support from the US NSF. D.L. was supported by an Australian Research Council failure of two subalpine conifers. Ecology 99, 966–977 (2018). Laureate Fellowship. A.S.M. was supported by the Environment Research and Tech- 37. Pan, Y. et al. Age structure and disturbance legacy of North American forests. nology Development Fund (S-14) of the Japanese Ministry of the Environment and by Biogeosciences 8, 715–732 (2011). the Grants-in-Aid for Scientific Research of the Japan Society for the Promotion of 38. Schurman, J. S. et al. Large-scale disturbance legacies and the climate Science (15KK0022). G.L.W.P. acknowledges support from a Royal Society of New sensitivity of primary Picea abies forests. Glob. Chang. Biol. https://doi.org/ Zealand Marsden Fund grant. S.L.S. acknowledges funds from the US Joint Fire Sciences 10.1111/gcb.14041 (2018). Program (project number 14-1-06-22) and UC ANR competitive grants. M.S. and T.H. 39. Gray, C. L. et al. Local biodiversity is higher inside than outside terrestrial acknowledges support from the institutional project MSMT CZ.02.1.01/0.0/0.0/16_019/ protected areas worldwide. Nat. Commun. 7, 12306 (2016). 0000803. M.G.T. acknowledges funding from the University of Wisconsin-Madison Vilas 40. Buma, B. & Barrett, T. M. Spatial and topographic trends in forest expansion Trust and the US Joint Fire Science Program (project numbers 09-1-06-3, 12-3-01-3, and and biomass change, from regional to local scales. Glob. Chang. Biol. 21, 16-3-01-4). The study used data from the TRY initiative on plant traits (http://www.try- 3445–3454 (2015). db.org). The TRY initiative and database is hosted, developed and maintained by J. 41. Borrelli, P., Panagos, P., Langhammer, J., Apostol, B. & Schütt, B. Assessment Kattge and G. Boenisch (Max Planck Institute for Biogeochemistry, Jena, Germany). TRY of the cover changes and the soil loss potential in European forestland: first is currently supported by Future Earth/bioDISCOVERY and the German Centre for approach to derive indicators to capture the ecological impacts on soil-related Integrative Biodiversity Research (iDiv) Halle-Jena-Leipzig. forest ecosystems. Ecol. Indic. 60, 1208–1220 (2016). ‘ fi 42. Westerling, A. L. Correction to Increasing western US forest wild re activity: Author contributions sensitivity to changes in the timing of spring’: Table 3. Philos. Trans. R. Soc. B Biol. Sci. 371, 20160373 (2016). A.S., C.S., and R.S. designed the study, analyzed the data and wrote the paper. B.B., A.W. 43. Seidl, R. et al. Modelling natural disturbances in forest ecosystems: a review. D., T.D., I.D.-H., S.F., L.E.F., Á.G.G., S.J.H., B.J.H., H.S.H., T.H., A.H., T.K., D.K., D.L., A. Ecol. Modell. 222, 903–924 (2011). S.M., J.M., J.P., G.L.W.P, S.L.S., M.S., M.G.T., and T.T.V. contributed data and com- 44. Anderegg, W. R. L. et al. Tree mortality from drought, insects, and their mented on the manuscript. interactions in a changing climate. New Phytol. 208, 674–683 (2015). 45. Littell, J. S., Mckenzie, D., Peterson, D. L. & Westerling, A. L. Climate and Additional information wildfire area burned in western U.S. ecoprovinces, 1916-2003. Ecol. Appl. 19, Supplementary Information accompanies this paper at https://doi.org/10.1038/s41467- 1003–1021 (2009). 018-06788-9. 46. Westerling, A. L., Turner, M. G., Smithwick, E. A. H., Romme, W. H. & Ryan, M. G. Continued warming could transform Greater Yellowstone fire regimes Competing interests: The authors declare no competing interests. by mid-21st century. Proc. Natl Acad. Sci. USA 108, 13165–13170 (2011). 47. Olson, D. M. et al. Terrestrial of the world: a new map of life on Reprints and permission information is available online at http://npg.nature.com/ Earth. Bioscience 51, 933 (2001). reprintsandpermissions/ 48. Fraley, C., Raftery & Adrian, E. Model-based methods of classification: using the mclust software in chemometrics. J. Stat. Softw. 18,1–13 (2007). 49. Kattge, J. et al. TRY—a global database of plant traits. Glob. Chang. Biol. 17, 2905–2935 (2011). Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in fi 50. Hothorn, T., Hornik, K., Wiel, M. Avande & Zeileis, A. Implementing a class published maps and institutional af liations. of permutation tests: the coin package. J. Stat. Softw. 28,1–23 (2008). 51. R Core Team. R: A Language and Environment for Statistical Computing. (RC Team, Vienna, Austria, 2018). 52. Bolker, B. M. et al. Generalized linear mixed models: a practical guide for Open Access This article is licensed under a Creative Commons ecology and evolution. Trends Ecol. Evol. 24, 127–135 (2009). Attribution 4.0 International License, which permits use, sharing, 53. Riley, S. J., DeGloria, S. D. & Elliot, R. A Terrain ruggedness index that , distribution and reproduction in any medium or format, as long as you give qauntifies topographic heterogeneity. Intermt. J. Sci. 5,23–27 (1999). appropriate credit to the original author(s) and the source, provide a link to the Creative 54. Grechka, D. A. et al. Universal, easy access to geotemporal information: Commons license, and indicate if changes were made. The images or other third party FetchClimate. Ecography 39, 904–911 (2016). material in this article are included in the article’s Creative Commons license, unless 55. Akaike, H. in Proc. 2nd International Symposium on Information Theory (eds indicated otherwise in a credit line to the material. If material is not included in the Petrov, B. N. & Caski, F.) 199–213 (Akadimiai Kiado, Budapest, 1998). article’s Creative Commons license and your intended use is not permitted by statutory 56. Marini, L. et al. Climate drivers of bark beetle outbreak dynamics in Norway regulation or exceeds the permitted use, you will need to obtain permission directly from spruce forests. Ecography 40, 1426–1435 (2017). the copyright holder. To view a copy of this license, visit http://creativecommons.org/ 57. King, G. & Zeng, L. in Encyclopedia of Biopharmaceutical Statistics (ed Chow, licenses/by/4.0/. S.-C.) (Marcel Dekker, New York, NY, 2004). 58. Gelman, A. & Hill, J. Data Analysis Using Regression and Multilevel/ Hierarchical Models (Cambridge University Press, Cambridge, MA, 2006) © The Author(s) 2018 59. Bates, D., Mächler, M., Bolker, B. & Walker, S. Fitting linear mixed-effects models using lme4. J. Stat. Softw. 67,1–48 (2014).

NATURE COMMUNICATIONS | (2018) 9:4355 | DOI: 10.1038/s41467-018-06788-9 | www.nature.com/naturecommunications 9